基于幅值谱与不变矩的特征提取方法及应用

Feature Extraction Based on Amplitude Spectrum and Moment Invariants and Its Application

研究了傅立叶变换、不变矩的原理及特点,提出基于幅值谱与不变矩的特征提取方法,并应用于中厚板的表面缺陷自动分类．从现场在线采集中厚板的表面图像,将每幅表面图像划分成128×128大小的子图像,对子图像进行傅立叶变换得到子图像的幅值谱,再对幅值谱图像求Hu不变矩,将不变矩作为特征量,通过这种方法提取的特征向量不仅具有平移、旋转不变性,并且具有抗噪、抑制光照不均的优点．将本文方法得到的特征量作为基于LVQ神经网络的分类器输入,对缺陷样本进行学习和分类,结果表明,这些特征量适用于中厚板表面缺陷的分类,识别率达81．5%．

Principles and characters of Fourier transform and moment invariants were studied. Feature extraction based on amplitude spectrum and moment invariants was proposed and applied to classification of surface defects of medium and heavy plates. Surface images captured from a production line of heavy and medium plates were divided into sub-images of size 128× 128. Amplitude spectrums of sub-images were obtained with fast Fourier transform. Hu Moment Invariants of amplitude spectrums were calculated as features, which were not only translation invariant and rotation invariant but also insensitive to noises and uneven illumination. The features were fed into classifiers based on LVQ networks. Results of training and testing with samples of defects showed that the features were applicable to classfication of surface defects of medium and heavy plates, and the classification rate was 81.5%.